Anthropic Launches Claude Managed Agents in Public Beta
Anthropic launched Claude Managed Agents on April 8, 2026, a cloud service that handles the infrastructure required to build, deploy, and run autonomous AI agents without developers managing containers, state, or orchestration themselves. The service entered public beta and is available to all Anthropic API accounts.
The core offering is a hosted execution layer: developers describe the tasks an agent should perform, specify which third-party tools it should access, and define security rules governing tool usage. Claude Managed Agents automatically provisions isolated containers for each agent, handles state tracking across long-running sessions, orchestrates tool calls, and performs error recovery. The platform also includes a multi-agent capability in research preview, allowing agents to spawn sub-agents for complex parallel tasks, and automatic prompt refinement that Anthropic says improved task success rates by up to 10 percentage points in internal testing.
Pricing is additive: $0.08 per session hour on top of standard Claude API token costs. Early adopters include Notion, Rakuten, and Asana, with Anthropic claiming the service shortens agent development timelines from months to weeks by eliminating boilerplate infrastructure. The service is accessible through the Anthropic Console, Claude Code, and a dedicated CLI. This positions Anthropic squarely in competition with similar hosted-agent offerings from AWS Bedrock Agents and Google Cloud's Vertex AI Agent Engine.
Read more — SiliconANGLE
The New Stack — The New Stack
Claude Code v2.1.98–v2.1.101: Vertex AI Wizard, Team Onboarding, and CA Certificate Improvements
Claude Code shipped four versions between early April 8 and April 10, 2026, with a cluster of features aimed at enterprise deployment, team ramp-up, and reliability.
The flagship addition in v2.1.98 (released April 9) is a Google Vertex AI setup wizard accessible from the login screen when selecting "3rd-party platform." The wizard walks through GCP authentication, project and region selection, credential verification, and model pinning in an interactive terminal flow — the same pattern as the AWS Bedrock wizard that shipped in v2.1.92. Together, these cover both major cloud AI platforms with guided setup, removing a common friction point for enterprise teams deploying Claude Code in airgapped or compliance-constrained environments.
v2.1.101 introduced the /team-onboarding command, which generates a personalised ramp-up guide for new teammates based on your local Claude Code session history — a practical shortcut for getting a new hire productive with the specific patterns and workflows your team has developed. The /ultraplan command and other remote-session features now auto-create a default cloud environment instead of requiring web-based setup first.
On the reliability and enterprise front, OS CA certificate store trust is now enabled by default for enterprise TLS proxies (override with CLAUDE_CODE_CERT_STORE=bundled). Rate-limit error messages now specify which limit was hit and when it resets. Refusal errors include API-provided explanations, and tool-not-available errors explain the reason and suggest remediation. A command injection vulnerability in the POSIX which fallback for LSP binary detection was also patched in this release cycle.
Read more — Claude Code Changelog
Transformers.js v4: WebGPU Runtime and Support for 20B+ Parameter Models in the Browser
Hugging Face released Transformers.js v4 on February 9, 2026, with the most significant architectural change since the library launched: a complete rewrite of the inference runtime in C++ in collaboration with the ONNX Runtime team, replacing the previous JavaScript/WebAssembly backend with a WebGPU-accelerated engine.
The practical developer impact is substantial. BERT-based models run approximately four times faster via the new com.microsoft.MultiHeadAttention operator. Build times dropped from 2 seconds to 200ms (a 10× improvement), and the total bundle size shrank by roughly 10%, with the web-only bundle (transformers.web.js) cut by 53%. The new runtime also unlocks support for models larger than 8 billion parameters: GPT-OSS 20B runs at approximately 60 tokens per second in benchmarks. Prior to v4, models above a few billion parameters were impractical in browser or edge environments.
New model architectures added include GPT-OSS, Chatterbox (voice generation), GraniteMoeHybrid, LFM2-MoE, and several others, alongside support for advanced patterns like Mamba state-space models, Multi-head Latent Attention, and Mixture of Experts. The ModelRegistry API reaches production-ready status with methods to check cache status and calculate download sizes before committing to a model pull. A companion Tokenizers.js standalone library ships as a separate 8.8kB (gzipped), zero-dependency package for environments that only need tokenization without the full inference stack.
The build system migrated from Webpack to esbuild, and the repository converted to a pnpm monorepo. The models.js file (previously 8,000+ lines) was split into modular files, making it significantly more maintainable and tree-shakeable.
Read more — Hugging Face Blog
Links & Sources
- Anthropic launches Claude Managed Agents to speed up AI agent development — SiliconANGLE
- With Claude Managed Agents, Anthropic wants to run your AI agents for you — The New Stack
- Claude Code Changelog — GitHub
- Transformers.js v4: Now Available on NPM! — Hugging Face Blog
You May Also Enjoy
